ECCV 2026 论文列表

Multimodal large language models (MLLMs) have becomethe dominant paradigm for chart understanding, yet they suffer from acritical vulnerability: sycophancy. When users embed false premises inqueries about charts, models override visual evidence to comply with theuser’s misconception. Existing mitigations incur a severe safety-utilitytrade-off : suppressing sycophantic compliance inevitably triggers over-refusal of valid queries, degrading general performance. To address thischallenge, we introduce Re-Check, a training-free inference frameworkthat follows a “Verify-then-Answer” workflow: it first decomposes userqueries into atomic claims, then verifies each claim against the chartimage, and finally routes the generation through an adaptive three-way mechanism. To ensure verification reliability, we propose the Con-trastive Visual Dependency Score (CVDS), an information-theoreticmetric that measures the KL divergence between the model’s predic-tions with and without visual input, filtering out textual-bias-drivenmisjudgments. Through extensive experiments on diverse benchmarks,Re-Check improves overall accuracy by 15.01% (from 57.49% to 72.50%with Qwen3-VL-8B) on sycophancy scenarios, while maintaining compet-itive performance on general benchmarks such as ChartQAPro. Theseresults demonstrate a superior safety-utility trade-off. Code is availableat: https://github.com/X1Wang/ReCheck-code.

Noisy labels pose a significant challenge in training deep neu-ral networks, as corrupted annotations inevitably induce overfitting andsignificantly degrade generalization. However, existing methods typicallyrely on static metrics and overlook training dynamics, struggling to effec-tively separate clean samples from noisy ones under complex and highnoise scenarios. In this paper, we propose a temporal attention-basedframework for learning with noisy labels. Specifically, we introduce an at-tention temporal consistency module for reliable sample selection, whichseparates samples by quantifying the evolutionary stability of visual at-tention during training. Then, to address the scarcity of clean samples inhigh-noise scenarios, we design an adaptive attention masks module tosynthesize images from clean and noisy samples, which employs differen-tial masking strategies to generate high-quality training samples. Finally,we present a mask-guided feature alignment module, which introducesfeature consistency regularization to facilitate the learning of robustrepresentations for noisy data. Extensive experiments demonstrate thatour method outperforms state-of-the-art approaches, achieving accuracygains of 1.6%/2.8% on CIFAR-10/100 under 90% symmetric noise, re-spectively. Code is available at https://github.com/iCAN-SZU/MSA

Di_x001B_usion models have achieved remarkable success in im-age and video generation, yet the high computational cost of iterativesampling remains a critical bottleneck for practical deployment. Featurecaching has emerged as a promising acceleration paradigm by reusingor predicting intermediate features across timesteps. However, existingtraining-free methods apply uniform prediction strategies that cannotadapt to the heterogeneous feature dynamics, causing signi_x001C_cant qualitydegradation under high acceleration ratios. We propose LinCa , a fea-ture caching framework based on learnable invertible networks. LinCadecomposes cached features into sub-components with distinct conti-nuity properties via a lightweight invertible network and applies dif-ferentiated prediction orders matched to each component. The strictinvertibility guarantees lossless reconstruction back to the original fea-ture space, forming a uni_x001C_ed Decompose-Predict-Reconstruct pipeline.By training separate predictors for di_x001B_erent models and timestep seg-ments, LinCa adapts to heterogeneous feature dynamics. Experiments onFLUX, Qwen-Image, and HunyuanVideo demonstrate that LinCa, withless than 0.2% additional parameters, signi_x001C_cantly outperforms exist-ing methods and maintains near-lossless quality at 5-7× speedup. Code:https://github.com/QHR69/LinCa.

Accurately segmenting objects without any manual anno-tations remains one of the core challenges in computer vision. In thiswork, we introduce Selfment, a fully self-supervised framework that seg-ments foreground objects directly from raw images without human labels,pretrained segmentation models, or any post-processing. Selfment firstconstructs patch-level affinity graphs from self-supervised features and ap-plies NCut to obtain an initial coarse foreground-background separation.We then introduce Iterative Patch Optimization (IPO), a feature-space refinement procedure that progressively enforces spatial coherenceand semantic consistency through iterative patch clustering. The refinedmasks are subsequently used as supervisory signals to train a lightweightsegmentation head with contrastive and region-consistency objectives,allowing the model to learn stable and transferable object representations.Despite its simplicity and complete absence of manual supervision, Self-ment sets new state-of-the-art (SoTA) results across multiple benchmarks.It achieves substantial improvements on Fmax over previous unsupervisedsaliency detection methods on ECSSD (+4.0\%), HKUIS (+4.6\%), andPASCAL-S (+5.7\%). Moreover, without any additional fine-tuning, Self-ment demonstrates remarkable zero-shot generalization to camouflagedobject detection tasks (e.g., .910 S_m on CHAMELEON and .792 \protect \mathcal {F}_{_x0008_eta }^{\omega }on CAMO), outperforming all existing unsupervised approaches andeven rivaling the SoTA fully supervised methods. Codes and weights areavailable at: https://geshang777.github.io/Selfment/.

Camera intrinsics play a vital role in recovering 3D structurefrom 2D video. However, most 3D algorithms assume that intrinsics re-main fixed throughout input video, an assumption that frequently failsfor real-world in-the-wild videos. Consequently, estimating per-frame in-trinsics from RGB images is critical for enabling 3D methods to operaterobustly on dynamic intrinsics videos. Previously, InFlux has contributedto this research direction by establishing the first real-world benchmarkwith per-frame ground truth intrinsics for dynamic intrinsics videos, en-abling systematic evaluation of algorithms for this task. Nevertheless,existing methods remain inaccurate due to two key obstacles: (i) train-ing data for this task is scarce and lacks diversity in camera intrinsics;and (ii) existing benchmarks, including InFlux, are limited in scene andcamera motion diversity, making it difficult to properly evaluate methodperformance. To address both gaps, we present InFlux++, consisting oftwo components. InFlux++ Synth is a large-scale procedurally generatedsynthetic video dataset with 441K+ annotated frames from 1841 high-resolution videos, providing accurate per-frame ground truth intrinsicsfor training dynamic intrinsics prediction models; a subset additionallyincludes per-frame camera pose, depth, and surface normals. The videosfeature rich intrinsics diversity through changes in camera zoom andfocus over time, as well as dynamic objects and realistic rendering ef-fects such as lens distortion and defocus blur. InFlux++ Real is a large-scale real-world benchmark that extends InFlux with 514K+ newly cap-tured frames across 334 high-resolution videos, spanning a wider rangeof scenes and camera motions. Finetuning existing intrinsics predictionmethods on InFlux++ Synth consistently improves focal length estima-tion across both InFlux++ Real and InFlux, suggesting that syntheticsupervision is a promising direction for RGB-based intrinsics prediction.For the dataset, benchmark, code, videos, submission instructions, andlive leaderboard, please visit https://influx.cs.princeton.edu/.

Diffusion models have shown promise in drag-style editing.Previous works mainly focus on point-based drag, which is inherentlyambiguous. This paper focuses on region-based drag and introduces anovel In-Context Region-based Drag (ICRDrag) method. Under the in-context learning framework, ICRDrag consumes a source image, a sourceregion mask, and a target region mask, producing the target draggedimage. Built upon the basic in-context learning model, we introduce twonovel attention regularization: 1) image-mask attention consistency toensure that a target region attends to similar source regions for image andmask modalities; 2) source-target attention correspondence to ensure themutual correspondence between source and target regions. To facilitateregion-based drag, we also construct Paired Region Dataset (PRD), alarge-scale dataset with paired masks and images. Extensive experimentsshow that ICRDrag significantly outperforms existing methods in bothquantitative metrics and user studies, achieving superior editing accuracyand visual fidelity. The dataset, code, and model are available at https://github.com/bcmi/ICRDrag-Region-Drag-Editing.

Video-based world models have recently garnered increas-ing attention for their ability to synthesize diverse and dynamic vi-sual environments. In this paper, we focus on shared world modeling,where a model generates multiple videos from a set of input images,each representing the same underlying world in di!erent camera poses.We propose IC-World, a novel generation framework, enabling paral-lel generation for all shared world input images via activating the in-herent in-context generation capability of large video models. We fur-ther finetune IC-World via reinforcement learning, Group Relative Pol-icy Optimization, together with two proposed novel reward models toenforce scene-level geometry consistency and object-level motion consis-tency among the set of generated videos. Extensive experiments demon-strate that IC-World substantially outperforms state-of-the-art methodsin both geometry and motion consistency. To the best of our knowledge,this is the first work to systematically explore the shared world modelingproblem with video-based world models.

Vision-Language Models (VLMs) empower embodied agentsto execute complex instructions, yet they remain vulnerable to contex-tual safety risks where benign commands become hazardous due to subtleenvironmental states. Existing safeguards often prove inadequate. Rule-based methods lack scalability in object-dense scenes, whereas model-based approaches relying on prompt engineering suffer from unfocusedperception, resulting in missed risks or hallucinations. To address this,we propose an architecture-agnostic safeguard featuring Context-GuidedChain-of-Thought (CG-CoT). This mechanism decomposes risk assess-ment into active perception that sequentially anchors attention to inter-action targets and relevant spatial neighborhoods, followed by semanticjudgment based on this visual evidence. We support this approach witha curated grounding dataset and a two-stage training strategy utilizingReinforcement Fine-Tuning (RFT) with process rewards to enforce pre-cise intermediate grounding. Experiments demonstrate that our modelsignificantly enhances safety, improving risk match rates by over 30%compared to base models while reducing oversafety. Beyond hazard de-tection, the generated visual anchors serve as actionable spatial con-straints for downstream planners, facilitating explicit collision avoidanceand safety trajectory generation.

Diffusion-based methods have dominated the HOI genera-tion, as they enable critical contact fusions or signals to guide the dif-fusion process. However, they often result in high artifacts and unstableinteraction quality due to error accumulation during iterative denois-ing. In this work, we propose HOIMask, the first generative maskedframework for modeling HOI motion in discrete space. HOIMask firstencodes both motion sequences and contact-aware signals into discrete2D human and object token maps via HOI Vector Quantization (VQ),preserving fine-grained spatial-temporal structure beyond conventional1D representations. On this basis, a generative masked modeling frame-work is employed to jointly capture human-object interaction dynamics,leveraging a transformer architecture designed to model complex spatial-temporal and interaction dependencies. To generate more coherent andphysically plausible motions, we further introduce a novel contact-awarereconstruction guidance in discrete space during inference, which fusescontact signals to optimize HOI tokens that forces the generated mo-tion with higher spatio-temporal consistency. With craftily designed mo-tion interaction tokens, dedicated architecture and guidance strategy,HOIMask outperforms state-of-the-art diffusion-based methods, gener-ating more realistic and semantically aligned HOI motions. Please referto our project page for more results.

Multimodal Large Language Models (MLLMs) suffer fromsevere memory bottlenecks in streaming video understanding due to ac-cumulated Key-Value (KV) cache with the growth of video length. Exist-ing training-free compression methods employ the same strategy acrossall attention heads in one layer, but ignore their diverse functional roles intemporal representation. In this paper, we reveal the inherent temporalheterogeneity of attention heads in MLLMs through systematic Spatio-temporal Attention Profiling (SAP), and discover that heads naturallyspecialize into instantaneous, short-term, and episodic roles agnostic toinput video frames and their queries. Motivated by this finding, we pro-pose the Head-aware Heterogeneous Streaming Video Cache (H2 SVC),a training-free KV cache compression framework for efficient streamingvideo understanding. H2 SVC features an offline routing strategy that ex-ploits the distinct temporal behavior of heads in each layer and adapttheir KV cache to dedicated memory banks tailored for varying temporalreceptive fields. Furthermore, we develop Semantic Trajectory CurvatureEviction (STCE) for the long-term episodic bank to enhance semanticdiversity under strict streaming budget. STCE models the Value statesas a geometric trajectory to continuously evict predictable frames onsmooth paths, while preserving sharp semantic inflection points. Exten-sive experiments on widely adopted streaming benchmarks demonstratethat H2 SVC achieves state-of-the-art performance, and is comparableor superior to full-context offline models with significantly reduced KVcache memory and accelerated inference speed.

Merging multiple 3D Gaussian Splatting (3DGS) scenes intoa single unified Gaussian representation is essential for large-scale 3Dmapping and long-term map management. Despite its importance, thisarea remains underexplored, and existing solutions exhibit several limi-tations. Learning-based methods attempt direct correspondence betweenGaussian primitives and require training on large 3DGS datasets. Image-based optimization methods depend heavily on coarse initialization fromgeneric foundation models and often incur expensive refinement. Wepresent Graph-GSReg. Our method constructs a 3D scene graph froma 3DGS and its rendered images, reformulating 3DGS registration asa graph registration problem. The proposed 3D scene graph representseach 3DGS at a higher-level representation, enabling a globally consis-tent understanding of semantic information and structural context foraccurate registration. To further construct a seamless unified scene, weintroduce a Self-Supervised Test-Time Optimization. Naively mergingtwo 3D Gaussian scenes often suffers from occlusion artifacts such ashollows and floaters. To alleviate this issue, we refine the merged Gaus-sians to preserve visual consistency between the original scenes and themerged scene. We evaluate our method on real and synthetic bench-marks, demonstrating competitive registration accuracy and merged scenerendering quality.

Adversarial Robustness Distillation (ARD) aims to trans-fer both natural and robust knowledge from a strong teacher modelto a lightweight student model, thereby achieving robust performancein resource-constrained settings. The dual-teacher robustness distilla-tion paradigm further provides an inspiring solution for optimizing the‘accuracy-robustness’ trade-off. However, existing dual-teacher distilla-tion frameworks fail to fully align the supervision provided by teachergroup, resulting in suboptimal performance. To fully utilize the poten-tial of dual-teacher robustness distillation, we propose a novel multi-teacher framework called global consistency multi-teacher robustness dis-tillation (GCMRD), which improves the existing dual-teacher paradigmfrom three perspectives. First, for internal maximization, we introducean integrated soft-label attack objective to replace hard-label or single-teacher-guided objective, leading to consistent attack preferences acrossboth natural and robust teachers. Second, for external minimization, wedesign a fully-repulsion regularization mechanism based on the erroneousadversarial outputs of the natural teacher, which enables more compre-hensive utilization of teacher supervision. Finally, we employ the outputsof the natural teacher to guide the fine-tuning of the robust teacher, in-directly balancing the student model’s accuracy-robustness trade-off byenforcing consistency with the teacher ensemble. Experimental resultsdemonstrate that our GCMRD achieves state-of-the-art performance inboth natural accuracy and adversarial robustness.

Existing optical flow estimation methods broadly follow twoparadigms: iterative optimization and diffusion-based estimation. Iter-ative methods, exemplified by RAFT, achieve accurate flow estimationthrough recurrent refinement, but can still be challenged by large displace-ments and complex motion patterns. Diffusion-based methods introducegenerative modeling into optical flow and have shown promising resultsin such ambiguous regions. However, existing diffusion-based flow modelsusually denoise the entire dense flow field from Gaussian noise, includingsimple regions where reliable motion structure can already be estimatedby a lightweight network. This increases the denoising burden and maylead to slow convergence and unstable training. To address these issues,we introduce FlowPainter, a diffusion-based optical flow framework thatreformulates dense-flow generation as confidence-guided soft inpainting.FlowPainter first employs a lightweight confidence-aware network to pre-dict a rough flow and a pixel-wise confidence mask, which serves as areliability gate for distinguishing reliable simple regions from uncertainhard regions. The resulting simple-flow prior is used for confidence-basedinitialization and is further injected into the iterative denoising processthrough confidence-gated residual guidance. With a dynamically decayingguidance strength, FlowPainter stabilizes early denoising while preservingthe flexibility of the diffusion model for late-stage detail refinement. Ex-tensive experiments on public benchmarks, including Sintel, KITTI, andSpring, demonstrate that FlowPainter achieves strong accuracy undercomparable training settings and improves convergence efficiency overexisting diffusion-based optical flow methods, with notable gains on chal-lenging benchmark splits. Our approach provides a practical direction forintegrating reliable discriminative priors with diffusion-based refinementfor optical flow estimation.

Federated Learning (FL) allows multiple devices to collaboratively train machine learning models while keeping their raw data private. However, in real-world scenarios, FL often struggles with unstable optimization and slow convergence. This is largely caused by heterogeneous client data, where both the data distributions and dataset sizes vary significantly across devices. To address these challenges, we propose FedDO, an adaptive federated optimization framework guided by reinforcement learning. Instead of using predefined rules or binary selection, FedDO formulates client coordination as a continuous control problem. Specifically, we employ a sample-efficient Distributional Soft Actor-Critic with Three Refinements (DSAC-T) agent to dynamically allocate a continuous data-usage ratio for each client per round. This fine-grained adjustment effectively mitigates gradient drift. Furthermore, we introduce a low-rank parameterization technique to compress the agent’s action space, ensuring scalability to thousands of clients. Extensive experiments on standard image classification benchmarks show that FedDO consistently achieves higher accuracy and much faster convergence than state-of-the-art FL baselines, especially under highly unbalanced and heterogeneous data settings. Code is available at https: //github.com/leafuan/FedDO_code.

Image-based Virtual Try-On (IVTON) has greatly advancedthrough diffusion models, yet existing methods require many samplingsteps and depend on masks with costly auxiliary networks. In addition,the absence of large-scale mask-free paired datasets further limits thedevelopment of mask-free IVTON. We propose FDM-MFVT, a few-stepdiffusion model for mask-free IVTON, integrating an Outfit-aware NoiseOptimization Module (OANO) and an Instruction-driven Try-on Module(IDT) to enhance efficiency and flexibility.The OANO module initializesthe alignment space with noise using the input image and only needs 6steps to generate a higher-fidelity try-on image compared to 30 steps.TheIDT module uses virtual try-on prompts and efficient adaptation to gen-erate high-quality results from garment and person images alone. Wefurther introduce MFVT, a 30,000-pair mask-free IVTON dataset. Ex-periments show that FDM-MFVT achieves superior quantitative andqualitative results with fewer inference steps than mask-based and mask-free baseline methods.

Vision-language models (VLMs) have achieved strong per-formance across diverse multimodal tasks, yet they remain vulnerable tounreliable reasoning. Existing self-correction methods mitigate these is-sues but typically rely on post-training or carefully engineered feedback,incurring high computational cost. In this work, we revisit this challengethrough the lens of emotional cues, asking whether they can activate la-tent self-correction behaviors in VLMs without additional training. Wefind that emotional signals serve as an effective trigger for self-correction, encouraging more cautious and reflective reason-ing. Motivated by this finding, we propose ESC (Emotional Self-Correction), a training-free self-correction framework. ESC introducesan external verifier that detects potentially incorrect initial responsesand injects emotional feedback to encourage model to reflect, and pro-duce a better revised response without additional training. Extensiveexperiments across safety, hallucination, vision-centric perception, andmultimodal reasoning benchmarks show that ESC consistently improvesreliability while preserving overall model utility. These results suggestthat emotion can function not only as an ability to be recognized, butalso as a practical control signal for scalable self-correction in VLMs. Wetherefore believe that ESC provides a strong foundation for anew reliable human-like, emotion-integrated research direction.Our project is publicly available at https://genai4e.github.io/ESC/.

Vision-language models (VLMs) have recently shown remark-able capabilities in visual understanding and generation, but remainvulnerable to adversarial manipulations of visual content. Prior object-hiding attacks primarily rely on suppressing or blocking region-specificrepresentations, often creating semantic gaps that inadvertently inducehallucination, where models invent plausible but incorrect objects. In thiswork, we demonstrate that hallucination arises not from object absenceper se, but from semantic discontinuity introduced by such suppression-based attacks. We propose a new class of background-consistent objectconcealment attacks, which hide target objects by re-encoding their vi-sual representations to be statistically and semantically consistent withsurrounding background regions. Crucially, our approach preserves tokenstructure and attention flow, avoiding representational voids that trig-ger hallucination. We present a pixel-level optimization framework thatenforces background-consistent re-encoding across multiple transformerlayers while preserving global scene semantics. Extensive experimentson state-of-the-art vision-language models show that our method effec-tively conceals target objects while preserving up to 86% of non-targetobjects and reducing grounded hallucination by up to 3× compared toattention-suppression-based attacks. Qualitative results further confirmthat our approach maintains scene coherence and avoids spurious ob-ject insertion. Our findings highlight semantic continuity as a key factorin hallucination behavior and introduce a new direction for adversarialanalysis of generative multimodal models.

Progress in dental computer vision is limited by the absence of large-scale multimodal datasets that jointly capture 3D intraoral geometry and 2Dappearance across diverse clinical settings. Existing resources are typically uni-modal, which hinders robust cross-modal learning and generalization. We as-semble and release a multi-center dataset of 1,000 patients comprising 2,000registered upper/lower intraoral scans, 5,000 paired intraoral photographs, and2,403 clinician-authored reports. This combination links detailed 3D dental ge-ometry with complementary 2D evidence, supporting occlusal and orthodonticanalysis. Moreover, to enable scalable and privacy-preserving acquisition andannotation across distributed centers, we introduce an open platform that sup-ports multimodal ingestion and structured labeling. Experiments indicate thatstate-of-the-art multimodal models fail to generate clinically faithful reports, moti-vating geometry-aware adaptation. We therefore propose IOS-Qwen, which fuses aPointTransformer 3D encoder with Qwen3-VL to generate structured, point-cloud-conditioned reports. Together, the dataset, the platform, and the baselines establisha foundation for multimodal dental AI research. Code is publicly released.3

Image generative models aim to sample data points fromthe underlying data manifold, a task that requires learning and decod-ing a dense, low-dimensional, and compact parameterization space. Toachieve this, we propose the Data Manifold-aware Image diffusioN moDel(MIND), a novel framework that explicitly models manifold geometry byintegrating discrete patch tokenization into the score function of a con-tinuous diffusion model. This approach successfully leverages both thestructural quantification capabilities of discrete tokens and the parallelgeneration flexibility of continuous diffusion. Moreover, we enable end-to-end differentiable training via a novel soft top-k aggregation mechanismand introduce dual-branch high-frequency feature embedding layers toalleviate the spectral bias of transformer backbones on low-dimensionalinputs. Furthermore, for inference, we design a multi-stage transitionsampling scheme that dynamically adjusts the sampling scheme based ontimestep. Extensive experiments on ImageNet 256×256 demonstrate theeffectiveness of MIND. After 80-epoch training, our base model achievesan FID of 22.73 without guidance, nearly halving the 43.47 FID of thevanilla DiT-B/2 baseline. The proposed method reduces FID by 15.95and 9.06 on average compared with the baselines DiT and SiT, respec-tively. For image generation on ImageNet 256×256 with guidance, theproposed MIND-B with only 130M parameters achieves an FID of 2.06,surpassing LlamaGen-3B with 3.1B parameters. Our MIND-XL with715M parameters further reduces the FID to 1.95. Our MIND intro-duces a fresh perspective on diffusion-based image generation, pavingthe way for future research and innovation in this community. The codeis available here: https://github.com/xddgit/MIND.